Blind separation of hybrid mixture signals based on kernel density maximum entropy approach
Hong Li · Jisuanji yingyong yanjiu · 2010
In some traditional methods of independent component analysis, the nonlinear evaluation functions are always chose empirically for demixing mixture signals.There is a dramatic degradation of the performance of these methods when the mixture contains super-Gaussian and sub-Gaussian ones at one time.Kernel density maximum entropy method(KD-MEM)is a parametric probability density estimated based on moments of random variables.In this paper,by use of kenel density maximum entropy approach(KD-MEM),directly evaluated score functions by probability density funcions of signals estimated straightly from observed data,and gave a stochastic gradient method to separate independent components.The algorithm proposed succeeds in separating the hybrid mixtures of source signals which included Gaussian, super-Gaussian and sub-Gaussian ones it is truly blind to the sources.Simulations confirm the effectiveness of the proposed algorithm.